Recursive launches Flint, a free coding model built by AI

Recursive, the startup that raised $650 million to automate AI research itself, released its first model on Wednesday: Flint, a free coding model that runs on a laptop and was built largely by the company's own AI research system.

Flint is a 40-billion-parameter mixture-of-experts model that activates only 2.5 billion parameters per task, which is what lets it fit on ordinary hardware instead of a GPU cluster. The weights are downloadable on Hugging Face under the Apache 2.0 license, so teams can run it offline, keep source code private, and pay nothing per token.

What happened

Recursive says its automated research system, the Eureka Machine, contributed to nearly every stage of Flint's development, from pretraining and post-training through inference work. The result: a small team shipped the model from scratch in under four months.

The performance numbers are what make this interesting rather than symbolic. On a MacBook Pro with an M4 Pro chip, Flint generated about 131 tokens per second in Recursive's tests, roughly double the 58 to 69 tokens per second that Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning managed on the same machine. On Nvidia's RTX Spark it cleared 150 tokens per second. On SWE-Bench Verified, which measures real bug-fixing, it beat every model under 50 billion parameters, and only Alibaba's Qwen3.8 27B scored higher on Recursive's own six-test average.

Why it matters

Two trends collide here. The first is local inference: a model that runs on a developer's own machine removes the per-token meter, the network round trip, and the compliance question of sending proprietary code to a vendor. The second is AI building AI. Flint is the first widely downloadable artifact produced mostly by an automated research loop, which makes it a public test of recursive self-improvement, one of the most debated ideas in the field.

Wired's Will Knight tested Flint on a Mac with 48GB of memory and found it nailed his coding requests every time, though Recursive itself admits the model needs more guidance than paid assistants and recommends breaking large tasks into smaller steps.

The builder takeaway

For teams wiring agents into production, Flint changes the cost model of the code-writing step. An agent that writes code can now call a local model for routine edits and reserve a frontier API for the hard cases. The catch is orchestration: smaller models degrade on long, multi-step tasks, so the routing layer that decides which model handles which subtask becomes the part that actually determines quality.

Recursive frames the release as safe, supervised self-improvement with humans in the loop. Whether that holds as the Eureka Machine takes on larger models is the open question the company has now put in front of everyone who downloads the weights.